Medical Image Reconstruction Combining Mathematical Equation Model And Deep Learning Algorithm

Medical imaging technologies such as CT and SPECT play an important role in the diagnosis of diseases.The imaging quality is not only related to the accuracy of medical equipment hardware,but also depends on the image reconstruction algorithm.Medical image reconstruction is susceptible to noise,and it is difficult to obtain accurate inversion results.In the field of image reconstruction,analytical method and iterative method are two main types of mathematical equation models.Analytical methods have fast calculation speed but poor imaging quality.Iterative methods have good imaging quality but have to take much time to calculate and cannot achieve realtime reconstruction.In recent years,deep learning algorithms have developed rapidly,especially convolutional neural networks,which have shown strong performance in image processing.In view of disadvantages of traditional image reconstruction algorithms of low imaging quality and long calculation time,considering advantages of deep learning algorithms,the traditional image reconstruction algorithm and deep learning algorithm are combined to overcome the disadvantage of low imaging quality and shorten the calculation time.In this paper,the U-Net convolutional neural network is selected and some modifications are made to make it suitable for image reconstruction tasks.The classic iterative algorithm is combined with the modified U-Net convolutional neural network to establish a joint network framework.The mathematical expression of the joint network and calculation methods of network parameters are presented.Using the classic algebraic reconstruction algorithm(ART)and the maximum likelihood-maximum expectation algorithm(ML-EM)as the mathematical equation model in the joint network framework,combined with deep learning algorithm,this paper proposes the ART U-Net joint network and ML-EM U-Net joint network.By constructing a data set,using projection data with noise as input and artificially synthesized real images as labels,the two joint networks are trained separately.We simulate two SPECT application scenarios such as lung perfusion imaging and myocardial perfusion imaging to test our methods.The trained ART U-Net joint network and the ML-EM U-Net joint network are used to reconstruct SPECT images.The image reconstruction effects of these two joint networks are analyzed from the aspects of image vision,MSE and PSNR indicators.Our methods are compared with the image reconstruction effects of the classic ART+TV and TV-EM algorithms.Results show that these two joint networks can reconstruct a SPECT image with no noise and artifacts and has a clear interface when the projection data contains varying degrees of noise.Effects on both MSE and PSNR are better than ART+ TV and TVEM algorithms.Moreover,these trained joint networks can reconstruct images in a short time,have the advantage of fast imaging speed.They can make up for the defect that iterative algorithms cannot reconstruct images in real time.In addition,the algorithm framework combining mathematical equation models and deep learning algorithms is also suitable for image reconstruction in other fields.